Praxis Maldevide

maldv

·

https://www.maldevide.com/

maldevide

AI & ML interests

ai alchemy

Recent Activity

liked a model about 13 hours ago

Qwen/Qwen2.5-32B

liked a model about 17 hours ago

meta-llama/Llama-3.1-70B

repliedto grimjim's post 1 day ago

I tinkered a bit with refusal ablation after seeing this article that proposed using optimal transport. https://arxiv.org/abs/2603.04355 Title: Efficient Refusal Ablation in LLM through Optimal Transport Authors: Geraldin Nanfack, Eugene Belilovsky, Elvis Dohmatob While closed Bures-Wasserstein sounded cool, the use of top-K at low K implied that the approach likely wouldn't capture refusal in detail, and increasing KL divergence with greater K suggested that the approach was too damaging for latent reasoning. I tried it out, and the results were disappointing.

View all activity

Organizations

models 47

maldv/Qwen3.6-27B-Isometry-Terse

Image-Text-to-Text • 28B • Updated Jul 31 • 830 • 3

maldv/Gemma-4-31B-Isometry-RP

Text Generation • 33B • Updated Jul 27 • 77 • 5

maldv/Gemma-4-31B-Isometry-Fabled-Persona

Text Generation • 31B • Updated Jul 27 • 81 • 6

maldv/Olmo-3.1-32B-ThinkInstruct

Text Generation • 32B • Updated Dec 13, 2025 • 10

maldv/QwentileLambda2.5-32B-Instruct

Text Generation • 33B • Updated Nov 30, 2025 • 18 • • 2

maldv/Shisutemu-Masuta-Q3-32B

Text Generation • 33B • Updated Aug 18, 2025 • 13 • 1

maldv/Eva-Mindlink-72b

Text Generation • 73B • Updated Aug 6, 2025 • 17 • 2

maldv/Doctor-Kunou-72b

Text Generation • 73B • Updated Jul 21, 2025 • 11 • 3

maldv/praxis-bookwriter-qwen2.5-14b-sft

Text Generation • 15B • Updated Jun 5, 2025 • 16 • • 2

maldv/praxis-bookwriter-r8-qwen2.5-14b-sft-lora

Updated Jun 4, 2025 • 5 • 1

datasets 1

maldv/crabcanon

Viewer • Updated Mar 7, 2024 • 79.2k • 57 • 1